Research edition · 14 September 2026
The report’s observations, prices and forecasts are a dated snapshot. Examples of savings are estimates unless explicitly identified as study results. This is not a fresh verification of every claim.
Foundation models and APIs
↓
Assistants and copilots ─────── Personal context / memory
↓ ↓
AI-native applications ─────── AI-enhanced productivity suites
↓ ↓
Workflow automation and agents ─ Systems of record and business apps
↓
Security, permissions, audit, evaluation, provenance, and human approval
Definitions
| Term | Core idea | Typical input | Typical output or action | Main failure mode |
|---|---|---|---|---|
| Chatbot | Answers a prompt in a conversational interface | User text, files, sometimes web pages | Text, code, analysis, or media | Confidently wrong answer; no durable workflow state |
| AI assistant | General helper that may use personal context and tools | Prompt plus permitted history, files, apps, or device context | Advice, drafts, searches, and sometimes actions | Context is incomplete, stale, or broader than the user realizes |
| Copilot | AI embedded beside a human task | Current document, code, email, CRM record, or meeting | Suggestions, edits, summaries, formulas, or next steps | Human accepts plausible output without checking |
| AI-native application | The product’s main value depends on AI behavior | Job-specific material, examples, preferences | A finished draft, design, transcript, plan, or analysis | Output quality is uneven and the workflow may not export cleanly |
| AI-enhanced software | Existing software with AI features added | Existing records and user instructions | Search, generation, classification, or prediction inside the product | AI is bolted on without better data, permissions, or workflow fit |
| Automation tool | Deterministic triggers and actions across apps | Events, fields, schedules, webhooks | Data movement, notifications, record updates | Brittle logic, duplicate actions, or unexpected costs |
| AI agent | Model chooses steps and calls tools to reach a goal | Goal, context, tools, policies, and constraints | Multi-step work with possible external actions | Prompt injection, excessive agency, drift, or unreviewed side effects |
| Autonomous workflow | A repeatable process runs with little or no per-run approval | Trigger plus structured data and policy | End-to-end execution | Silent failure, bad edge-case decisions, or cascading errors |
| Multi-agent system | Several specialized agents coordinate | Shared state, roles, and handoffs | Decomposition, debate, execution, or review | Cascading errors, cost growth, and hard-to-audit decisions |
| Knowledge-management tool | Captures, organizes, retrieves, and reuses information | Notes, documents, links, highlights, conversations | Search, summaries, connections, and memory | Garbage-in/garbage-out; retrieval misses important context |
| Workflow platform | Combines data, process, permissions, and automation | Structured records plus triggers and rules | Operational workflows and reporting | Tool sprawl or a low-code system no one owns |
| Personal-assistant app | Optimizes an individual’s tasks, time, information, or communication | Calendar, inbox, tasks, preferences, voice, location | Planning, reminders, scheduling, capture, and prioritization | Over-optimizes the calendar or creates notification pressure |
| Enterprise AI platform | Governed AI across company sources and applications | Permissioned enterprise data and policies | Search, answers, agents, analytics, and actions | Overshared source data, complex rollout, or unclear ROI |
| Vertical AI application | AI designed for a specific industry or job | Domain records, policy, and workflow artifacts | Domain-specific recommendations or actions | Regulatory exposure and false confidence in high-stakes work |
What makes a product genuinely useful?
An AI feature is materially more valuable when it has:
- A specific job to be done, not only a blank prompt box.
- Relevant context retrieved from the current task, source documents, or system of record.
- Structured outputs that can be checked or inserted into the next step.
- Action boundaries that make clear what the system may read, write, send, buy, delete, or change.
- Feedback and correction that improve the current result without silently training on sensitive data.
- Integration with the place where the work already lives.
- Observable quality, including citations, confidence, source links, diffs, logs, or evaluation scores.
- Predictable economics, including clear limits and a kill switch for overages.
The weakest version is a chatbot placed beside a static database. The strongest version is a grounded, permission-aware, reviewable workflow that reduces coordination cost without hiding judgment.